AI Engineering
The smallest AI that solves it.
Setup help, agent builds, custom models — from an engineer who'll tell you which one you actually need. Most people asking for a custom model don't need one.
Find the level your problem actually lives on.
Each one costs more, takes longer, and breaks in more ways than the one below it. The job of the first conversation is to find the lowest level that solves your problem — and stop there.
Get the tools working properly
The right AI tools for your team, configured for your actual work — shared workspaces, saved assistants for the jobs you repeat, prompt templates, and rules for what not to put in.
- Right when
- Your team is either not using AI or using it badly, and the work is judgement-heavy rather than data-heavy.
- Not when
- The answer lives in your documents, your database, or last week's numbers. No amount of prompting invents facts the model was never given.
Point it at your own information
Your documents, policies, product data, or history made searchable by the model — so answers come from your business instead of the internet's average.
- Right when
- The knowledge exists but is scattered, and it changes. Retrieval keeps answers current and lets you check where each one came from.
- Not when
- Nobody can say where the information lives or who owns it. That's a filing problem, and AI makes filing problems worse, not better.
Let it do the work, not just answer
Multi-step work that crosses systems — qualify the lead, draft the reply, update the CRM, book the slot — with hard limits on what it may read, prepare, do, and never touch.
- Right when
- The process is repetitive, already written down, and you can name what "wrong" looks like and who catches it.
- Not when
- The process only exists in someone's head. Automating an undocumented process just makes the confusion faster.
Train the model itself
Fine-tuning a model on your data so it behaves a specific way by default. Genuinely useful — for a narrow set of problems that is much smaller than the number of people asking for it.
- Right when
- You've already done the three below, still miss your accuracy or consistency target, and run enough volume to earn the cost back.
- Not when
- Almost always. The training run is cheap; preparing the data is most of the budget, and every question you ask the finished model costs several times more than the stock one — forever.
Answer ten questions. Get a straight recommendation.
No sales call to book a sales call. You write it down, I read it, and I tell you which level you're on — including when the answer is that you don't need me.
Tell me
Ten questions about your business and the work you want help with. About four minutes.
I read it
I reply within one business day with the level I think you're on and what it would take.
Session or Blueprint
A $250 working session if it's small. A $2,500 Blueprint if it's a real build.
Build & run
We build it, wire it into your systems, and keep it working once it's live.
Nothing you pay for is wasted going up. The session fee comes off the Blueprint. The Blueprint comes off a build booked within 60 days. And the Blueprint is yours to keep — including if it tells you not to build.
Three things most people selling this won’t tell you.
You probably don't need your own model.
It's the most requested thing and the least often correct. Prompting first, then your own documents, then agents. Training a model is the last resort, not the impressive first move.
The model is never the expensive part.
Connecting it to the systems you already run is roughly half the cost of any real project, and keeping it working costs a fifth of the build every year. Anyone quoting you on model choice alone is quoting the cheap half.
“Don’t build this” is a valid result.
A Blueprint that ends in a recommendation not to proceed has saved you the build. You keep the document either way, and you'll know exactly why.
Before you write it up.
Do I have to pay to find out what I need?
No. The ten questions are free and I answer them personally. You only pay once we both know what we're doing — a $250 session for something small, or a $2,500 Blueprint for a real build.
What if I already know I want an AI agent built?
Then say so in the questions and we'll skip ahead. I'll still want to know what it has to connect to, because that's where most of the cost and nearly all of the risk lives.
Can you actually build this, or do you just advise?
Both. Builds run from about $3,000 for a single workflow to $8,000–$25,000 for something spanning several systems. The advice and the build come from the same person, which is why the advice tends to be conservative.
What happens to my data?
Nothing leaves your control without you deciding it should. Where your data can and can't go is one of the first things the Blueprint settles — including whether anything needs to stay on your own infrastructure.
I have a bigger operations problem, not just AI.
Then start with a Scope Assessment instead — it’s the right front door for process, systems, and ERP work.
Ten questions. Four minutes. A straight answer.
Free, and I read every one myself. You'll hear back within one business day with the level you're on and what it would cost.
Tell me what you're working on